Human: and Assistant: are special terms that Claude has been trained to think of as indicators of who is speaking. This means you should never have a human message that gives examples of dialogue containing Human: and Assistant:.

Use H: and A: for examples

Consider the following prompt:



Human: I’m going to show you a sample dialogue and I want you to tell me if the response from the assistant is good. Here is the sample:

<sample_dialogue>
Human: What is your favorite color?
Assistant: I don’t have a favorite color.
</sample_dialogue>

What do you think of this dialogue?

Assistant:

You may think that the assistant will read this as a single message from the human just like we do, but the assistant will read the dialogue above as follows:

There was this message from a human to the assistant:
Human: I’m going to show you a sample dialogue and I want you to tell me if the response from the assistant is good. Here is the sample: <sample_dialogue>
Then there was this second message from the human to the assistant:
Human: What is your favorite color?
Then there was the following reply from the assistant to the human:
Assistant: I don’t have a favorite color. </sample_dialogue> What do you think of this dialogue?
And finally there was a prompt for the assistant to give another reply to the human:
Assistant:

This is very confusing to the assistant.

This is why, if you give examples of dialogue, you must replace Human: and Assistant: with something else, such as User: and AI: or H: and A:.

For example, the following edited version of the prompt above will work just fine:



Human: I’m going to show you a sample dialogue and I want you to tell me if the response from the assistant is good. Here is the sample:

<sample_dialogue>
H: What is your favorite color?
A: I don’t have a favorite color.
</sample_dialogue>

What do you think of this dialogue?

Assistant:

In this case the assistant sees a single message from the human that includes a sample dialogue, and it then sees a prompt for it to respond at the end, which is what we wanted.

Use Human: and Assistant: to put words in Claude's mouth

You should use Human: and Assistant: tokens in your prompt if you want to pass Claude a previous conversation. One way to get Claude to do something is to show it previously asking or agreeing to do so, like this:



Human: I have two pet cats. One of them is missing a leg. The other one has a normal number of legs for a cat to have. In total, how many legs do my cats have?

Assistant: Can I think step-by-step?

Human: Yes, please do.

Assistant:

In this case, you want Claude to think it actually asked to think step-by-step and you gave it permission to do so. Proper usage of the Human: and Assistant: tokens will accomplish this.